融合篇章上下文有效识别的篇章级机器翻译

Document-level machine translation fused effective context recognition

篇章翻译是近来兴起的研究热点,如何在翻译文档时充分利用篇章信息一直是该研究的关键点和难点.在篇章级机器翻译中,如何选取当前句的篇章上下文是非常关键的.虽然相关研究使用的篇章上下文不尽相同,但是却少有在选取之前对上下文信息进行识别筛选.本文提出了一种融合篇章上下文有效识别的篇章级翻译模型,引入判别篇章上下文是否有效的分类任务,并根据判别结果来控制目标端对篇章上下文的利用.在中英、英德翻译任务上,与基准系统相比,本模型的翻译性能都得到了显著的提升.

Recently,the document-level machine translation has been regarded as a research hotspot.However,how to make full use of document-level information when translating documents has continued to face difficulties.Similarly,how to select the context of the current sentence is very critical.Although selection methods in previous studies abound,however,it is rare to filter the context information before selection.This paper proposes a document-level translation model that incorporating effective context,introduces the classification task of identifying whether context is effective,and controls the use of context according to the identification results.Compared with the baseline system,our model has achieved strong improvement in Chinese-English and English-German translation tasks.